Figure 1.

Histogram of one AUC from true data vs. 1000 AUCs of classifiers built by the same greedy algorithm applied to pseudo data (NP and P labels randomly permuted). Fitted with a beta distribution, the AUC from real data indicates a P-value of 0.012. Since 17 random AUCs of 1000 by chance exceed the true AUC, an algebraic method³⁶ gives alternative P-value=0.018. Thus, the performance of the Greedy Algorithm limited to selection of at most six markers and applied to the full data set is unlikely to be chance. AUC, area under the curve of receiver operating characteristic.